LLM CLI Files

New features:

• llm logs --usage Markdown output now includes the response duration in milliseconds and as a human-readable duration. llm logs --short includes a new duration_ms field. #1653

Bug fixes:

• Significantly improved the performance of llm logs for long conversations by caching repeated message and model lookups. #1654
• Dynamically generated OpenAI options classes are now cached, avoiding repeated Pydantic class construction by plugins such as llm-openrouter. Thanks, waveplate. #1651
• Invalid schema DSL passed to llm prompt --schema now produces a clean command-line error instead of a Python traceback. Thanks, ikatyal2110. #1647
• llm --extract now recognizes fenced code blocks in responses that use CRLF line endings. Thanks, mameikagou. #1644
• monotonic_ulid() now remains monotonic if the system clock moves backwards or concurrent calls observe timestamps out of order. Thanks, Dylan Pulver. #1641
• Declared the typing-extensions package as a direct dependency, and added tests to protect against accidental missing dependencies. Thanks, Vansh Taneja. #1622